Back-translator

The back-translation module using fairseq translation models available in PyTorch Hub

Requirements

torch, nltk, tqdm, cython, fastBPE

Make sure to download the nltk's punkt tokenizer

$ python -c "import nltk; nltk.download('punkt')"

Example Usage

Load pretrained NMT models made by fairseq

src2tgt = "transformer.wmt19.en-de.single_model"
tgt2src = "transformer.wmt19.de-en.single_model" 
# Tokenizer and BPE differ between the models
tokenizer = "moses"
bpe = "fastbpe" 
device = "cuda"

bt = BackTranslator(src2tgt, tgt2src, tokenizer, bpe, device)

Backtranslate one sentence

sample_sentence = "Python is an interpreted, high-level, general-purpose programming language."
bt.backtranslate(sample_sentence)

# 'Python is an interpreted, sophisticated, universal programming language.'

You can also use multiple documents (The length of the documents does not have any limitation)

sample_doc = ['Python is dynamically typed and garbage-collected.', 
              'It supports multiple programming paradigms, including structured (particularly, procedural), object-oriented, and functional programming.', 
              'Python is often described as a "batteries included" language due to its comprehensive standard library']
bt.backtranslate_docs(sample_doc)

# ['Python is typed dynamically and garbage collected.',
# 'It supports several programming paradigms, including structured (especially procedural), object oriented and functional programming.',
# "Python is often described as the 'Battery Language' because of the comprehensive standard library."]

By using main.py, you can train your own tsv document (to see the detail format of the tsv file, see the example file)
(Notice that the tsv file does not have to contain label columns.)

# Example for using cpu
$ python main.py --data_dir=./data/imdb_sample.tsv --output_dir=./output/ --batch_size=32 

# for using single gpu (example for using only one gpu; gpu 1)
$ python main.py --data_dir=./data/imdb_sample.tsv --output_dir=./output/ --batch_size=64 --gpus 1

# for using multiple gpus (example for using two gpus; gpu 0 and 1)
$ python main.py --data_dir=./data/imdb_sample.tsv --output_dir=./output/ --batch_size=64 --gpus 0 1

Reference

google research UDA
fairseq NMT models